Improved object detection algorithm for drone-captured dataset based on yolov5

Abstract

Object detection is a basic task on computer vision, recently drone-captured scenarios had a wide range of applications in the industry. In this paper, firstly we will introduce some characteristics of yolov5s structure and some defects of yolov5s on tiny-size object detection. Secondly, to solve the defects mentioned above we will focus on the improvement based on the yolov5s structure to VisDrone dataset. Due to some characteristics of UAV dataset, the target scale changes greatly at different flight altitude, some target areas captured by drone show high density, which may bring some blurring or occlusion and some images may cover a large area, so we try to improve the structure of yolov5 for better performance. An additional prediction head is added to detect tiny-scale targets, which is also used to deal with large size variance of objects, and an EPSA net module is added in the middle of the backbone to explore the self-attention potential of feature representation. Moreover, extra up sampling is added in the neck part to optimize the detection of tiny-size target samples. Finally, it is proved by experiments that new yolov5s structure improve the value of [email protected] about 7% and improve the value of [email protected]:.95 about 5% on VisDrone-2019-DET dataset.